Message-ID: <1cf71906-ba99-e637-650f-fc08ac4f3d5f@redhat.com>
Date:   Wed, 6 Nov 2019 22:09:29 +0100
From:   Paolo Bonzini <pbonzini@...hat.com>
To:     Dan Williams <dan.j.williams@...el.com>,
        Sean Christopherson <sean.j.christopherson@...el.com>
Cc:     Radim Krčmář <rkrcmar@...hat.com>,
        Vitaly Kuznetsov <vkuznets@...hat.com>,
        Wanpeng Li <wanpengli@...cent.com>,
        Jim Mattson <jmattson@...gle.com>,
        Joerg Roedel <joro@...tes.org>, KVM list <kvm@...r.kernel.org>,
        Linux Kernel Mailing List <linux-kernel@...r.kernel.org>,
        Adam Borowski <kilobyte@...band.pl>,
        David Hildenbrand <david@...hat.com>
Subject: Re: [PATCH 1/2] KVM: MMU: Do not treat ZONE_DEVICE pages as being
 reserved

On 06/11/19 19:04, Dan Williams wrote:
> On Wed, Nov 6, 2019 at 9:07 AM Sean Christopherson
> <sean.j.christopherson@...el.com> wrote:
>>
>> Explicitly exempt ZONE_DEVICE pages from kvm_is_reserved_pfn() and
>> instead manually handle ZONE_DEVICE on a case-by-case basis.  For things
>> like page refcounts, KVM needs to treat ZONE_DEVICE pages like normal
>> pages, e.g. put pages grabbed via gup().  But KVM needs special handling
>> in other flows where ZONE_DEVICE pages lack the underlying machinery,
>> e.g. when setting accessed/dirty bits and shifting refcounts for
>> transparent huge pages.
>>
>> This fixes a hang reported by Adam Borowski[*] in dev_pagemap_cleanup()
>> when running a KVM guest backed with /dev/dax memory, as KVM straight up
>> doesn't put any references to ZONE_DEVICE pages acquired by gup().
>>
>> [*] http://lkml.kernel.org/r/20190919115547.GA17963@angband.pl
>>
>> Reported-by: Adam Borowski <kilobyte@...band.pl>
>> Debugged-by: David Hildenbrand <david@...hat.com>
>> Cc: Dan Williams <dan.j.williams@...el.com>
>> Cc: stable@...r.kernel.org
>> Signed-off-by: Sean Christopherson <sean.j.christopherson@...el.com>
>> ---
>>  arch/x86/kvm/mmu.c       |  8 ++++----
>>  include/linux/kvm_host.h |  1 +
>>  virt/kvm/kvm_main.c      | 19 +++++++++++++++----
>>  3 files changed, 20 insertions(+), 8 deletions(-)
>>
>> diff --git a/arch/x86/kvm/mmu.c b/arch/x86/kvm/mmu.c
>> index 24c23c66b226..bf82b1f2e834 100644
>> --- a/arch/x86/kvm/mmu.c
>> +++ b/arch/x86/kvm/mmu.c
>> @@ -3306,7 +3306,7 @@ static void transparent_hugepage_adjust(struct kvm_vcpu *vcpu,
>>          * here.
>>          */
>>         if (!is_error_noslot_pfn(pfn) && !kvm_is_reserved_pfn(pfn) &&
>> -           level == PT_PAGE_TABLE_LEVEL &&
>> +           !kvm_is_zone_device_pfn(pfn) && level == PT_PAGE_TABLE_LEVEL &&
>>             PageTransCompoundMap(pfn_to_page(pfn)) &&
>>             !mmu_gfn_lpage_is_disallowed(vcpu, gfn, PT_DIRECTORY_LEVEL)) {
>>                 unsigned long mask;
>> @@ -5914,9 +5914,9 @@ static bool kvm_mmu_zap_collapsible_spte(struct kvm *kvm,
>>                  * the guest, and the guest page table is using 4K page size
>>                  * mapping if the indirect sp has level = 1.
>>                  */
>> -               if (sp->role.direct &&
>> -                       !kvm_is_reserved_pfn(pfn) &&
>> -                       PageTransCompoundMap(pfn_to_page(pfn))) {
>> +               if (sp->role.direct && !kvm_is_reserved_pfn(pfn) &&
>> +                   !kvm_is_zone_device_pfn(pfn) &&
>> +                   PageTransCompoundMap(pfn_to_page(pfn))) {
>>                         pte_list_remove(rmap_head, sptep);
>>
>>                         if (kvm_available_flush_tlb_with_range())
>> diff --git a/include/linux/kvm_host.h b/include/linux/kvm_host.h
>> index a817e446c9aa..4ad1cd7d2d4d 100644
>> --- a/include/linux/kvm_host.h
>> +++ b/include/linux/kvm_host.h
>> @@ -966,6 +966,7 @@ int kvm_cpu_has_pending_timer(struct kvm_vcpu *vcpu);
>>  void kvm_vcpu_kick(struct kvm_vcpu *vcpu);
>>
>>  bool kvm_is_reserved_pfn(kvm_pfn_t pfn);
>> +bool kvm_is_zone_device_pfn(kvm_pfn_t pfn);
>>
>>  struct kvm_irq_ack_notifier {
>>         struct hlist_node link;
>> diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
>> index b8534c6b8cf6..0a781b1fb8f0 100644
>> --- a/virt/kvm/kvm_main.c
>> +++ b/virt/kvm/kvm_main.c
>> @@ -151,12 +151,23 @@ __weak int kvm_arch_mmu_notifier_invalidate_range(struct kvm *kvm,
>>
>>  bool kvm_is_reserved_pfn(kvm_pfn_t pfn)
>>  {
>> +       /*
>> +        * ZONE_DEVICE pages currently set PG_reserved, but from a refcounting
>> +        * perspective they are "normal" pages, albeit with slightly different
>> +        * usage rules.
>> +        */
>>         if (pfn_valid(pfn))
>> -               return PageReserved(pfn_to_page(pfn));
>> +               return PageReserved(pfn_to_page(pfn)) &&
>> +                      !is_zone_device_page(pfn_to_page(pfn));
> 
> This is racy unless you can be certain that the pfn and resulting page
> has already been pinned by get_user_pages().

What is the race exactly?

In general KVM does not use pfn's until after having gotten them from
get_user_pages (or follow_pfn for VM_IO | VM_PFNMAP vmas, for which
get_user_pages fails, but this is not an issue here).  It then creates
the page tables and releases the reference to the struct page.

Anything else happens _after_ the reference has been released, but still
from an mmu notifier; this is why KVM uses pfn_to_page quite pervasively.

If this is enough to avoid races, then I prefer Sean's patch.  If it is
racy, we need to fix kvm_set_pfn_accessed and kvm_set_pfn_dirty first,
and second at transparent_hugepage_adjust and kvm_mmu_zap_collapsible_spte:

- if accessed/dirty state need not be tracked properly for ZONE_DEVICE,
then I suppose David's patch is okay (though I'd like to have a big
comment explaining all the things that went on in these emails).  If
they need to work, however, Sean's patch 1 is the right thing to do.

- if we need Sean's patch 1, but it is racy to use is_zone_device_page,
we could first introduce a helper similar to kvm_is_hugepage_allowed()
from his patch 2, but using pfn_to_online_page() to filter out
ZONE_DEVICE pages.  This would cover both transparent_hugepage_adjust
and kvm_mmu_zap_collapsible_spte.

> This is why I told David
> to steer clear of adding more is_zone_device_page() usage, it's
> difficult to audit. Without an existing pin the metadata to determine
> whether a page is ZONE_DEVICE or not could be in the process of being
> torn down. Ideally KVM would pass around a struct { struct page *page,
> unsigned long pfn } tuple so it would not have to do this "recall
> context" dance on every pfn operation.

Unfortunately once KVM has created its own page tables, the struct page*
reference is lost, as the PFN is the only thing that is stored in there.

Paolo

Powered by blists - more mailing lists